Oct 25 3 days ago 04:19 v0.4.0: full-convergence checkpoints (S_safe = X + 2 epochs), TFLite exports, and updated evaluation archives Oct 22 6 days ago 02:50 v0.3.0: training-crop provenance (fixed T=180 crops for the released weights; step-0 loss fingerprint; training_provenance block; variable_length_training aligned with the weights), TFLite EER CIs and re-measured latencies, flaxspeaker>=0.3.0 quickstart Oct 20 8 days ago 18:27 Round 2: honest model card (training budget, variant labels, Vox1-O 10-spk subset, TFLite window, deviations), cleaned config.json, 26 dB Vox1-O subset metrics; remove misleading preprocessor_config.json 16:41 Use clean collection URL https://huggingface.co/collections/wq2012/flaxspeaker 16:09 Round 2: honest model card (training budget, variant labels, Vox1-O 10-spk subset, TFLite window, deviations), cleaned config.json, 26 dB Vox1-O subset metrics; remove misleading preprocessor_config.json 15:24 Round 2: honest model card (training budget, variant labels, Vox1-O 10-spk subset, TFLite window, deviations), cleaned config.json, 26 dB Vox1-O subset metrics; remove misleading preprocessor_config.json Oct 18 10 days ago 03:23 Round 2 sync: portable config.json and official VoxCeleb1-O eval_results.json 02:56 Round 2: Add scoring_type/loss_type/vad_config to config.json and voxceleb1_o_official to eval_results.json 00:22 Update ls_conformer_small_ge2e_contrast with non-saturating GE2E-Contrast loss and verified multi-domain evaluation Oct 15 13 days ago 01:38 Upload FlaxSpeaker pretrained model ls_conformer_small_ge2e_contrast 01:38 initial commit 
